Maintaining Discrimination and Fairness in Class Incremental Learning

By Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, Shutao Xia. 2019

560 citations recorded by the source corpus.

Deep neural networks (DNNs) have been applied in class incremental learning, which aims to solve common real-world problems of learning new classes continually. One drawback of standard DNNs is that they are prone to catastrophic forgetting. Knowledge distillation (KD) is a commonly used technique to alleviate this problem. In this paper, we demonstrate it can indeed help the model to output more discriminative results within old classes. However, it cannot alleviate the problem that the model tends to classify objects into new classes, causing the positive effect of KD to be hidden and limited. We observed that an important factor causing catastrophic forgetting is that the weights in the last fully connected (FC) layer are highly biased in class incremental learning. In this paper, we propose a simple and effective solution motivated by the aforementioned observations to address catastrophic forgetting. Firstly, we utilize KD to maintain the discrimination within old classes. Then, to further maintain the fairness between old classes and new classes, we propose Weight Aligning (WA) that corrects the biased weights in the FC layer after normal training process. Unlike previous work, WA does not require any extra parameters or a validation set in advance, as it utilizes the information provided by the biased weights themselves. The proposed method is evaluated on ImageNet-1000, ImageNet-100, and CIFAR-100 under various settings. Experimental results show that the proposed method can effectively alleviate catastrophic forgetting and significantly outperform state-of-the-art methods.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Radiology, Nuclear Medicine and Imaging, Meta & Continual Learning, AI Ethics, Fairness & Society

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Radiology, Nuclear Medicine and Imaging

In Meta & Continual Learning

In AI Ethics, Fairness & Society

Citations in the corpus

A Comprehensive Survey of Continual Learning: Theory, Method and Application (2023, background): “Label Space LwF [252], iCaRL [356], GEM [277], A-GEM [65], EEIL [53], BiC [457], WA [500], DER++ [46], ScaIL [33], S&B [218], SS-IL [11], Coil [506]”
S3C: Self-Supervised Stochastic Classifiers for Few-Shot Class-Incremental Learning (2023, methodology): “Several state-of-art-works [38,45,2,12,3] proposed different techniques to address the class imbalance problem in CIL like rescaling scores or balanced finetuning of classifiers, etc.”
First Session Adaptation: A Strong Replay-Free Baseline for Class-Incremental Learning (2023, background): “Thus, recent works either focus on fixing the softmax classifier [4, 1, 44] or deploy nearest class mean classifiers (NCM) as an alternative [25, 5, 11].”
RMM: Reinforced Memory Management for Class-Incremental Learning (2023, background): “…are either task-incremental learning—each phase corresponds to a task (dataset) that contains new data of all seen classes [7, 10, 19, 29, 43, 47, 57], or class-incremental learning (CIL)—each phase contains data of a new set of classes, i.e., classes are unseen [4, 6, 18, 25, 31, 37, 40, 41,…”
FeTrIL: Feature Translation for Exemplar-Free Class-Incremental Learning (2022, background): “Many recent solutions [4, 13, 33, 44, 46], based on deep nets, use replay from a bounded memory of the past to reduce forgetting.”
Generative Negative Text Replay for Continual Vision-Language Pretraining (2022, methodology): “In particular, those methods [52,47] aiming to alleviate class imbalance are not applicable to the continual learning of VLP models as there are no classlevel supervision for VLP.”
Anti-Retroactive Interference for Lifelong Learning (2022, methodology): “In addition, the distillation loss is implemented by [22,4,36] to reduce forgetting.”
Learning to Imagine: Diversify Memory for Incremental Learning using Unlabeled Data (2022, methodology): “Ldist is the widely-used distillation loss in previous works [4, 11, 17, 38], which forces the current feature space close to the old feature space to overcome forgetting and is formulated as:”
FOSTER: Feature Boosting and Compression for Class-Incremental Learning (2022, methodology): “Bic [38], WA [43] propose effective strategies to alleviate the bias of the classifier caused by imbalanced training data after distillation.”
General Incremental Learning with Domain-aware Categorical Representations (2022, neutral): “To deal with this problem, most works [4, 13, 34, 41] adjust the classifier weights or prediction logits between classes to eliminate the bias after learning the representations.”
Class-Incremental Learning with Strong Pre-trained Models (2022, methodology): “Class-incremental learning (CIL) methods [5,6,8,11,23, 29,34] learn to categorize more and more classes over time.”
Class-Incremental Learning by Knowledge Distillation with Adaptive Feature Consolidation (2022, methodology): “As the teacher’s knowledge, iCaRL [37] employs the sigmoid output of each class in the previous tasks while [9,48, 52] use the normalized softmax outputs with a temperature scaling.”
Few-Shot Class-Incremental Learning by Sampling Multi-Phase Tasks (2022, background): “Other works consider saving embeddings instead of raw images [38], using generative models for data rehearsal [39], task-wise adaptation [40], and output normalization [18], [41], [42].”
Constrained Few-shot Class-incremental Learning (2022, methodology): “In a similar vein, a weight aligning method has been proposed to align the norms of the weight vectors for novel classes to those for past classes [57].”
Forward Compatible Few-Shot Class-Incremental Learning (2022, methodology): “Other methods rehearsal former instances to overcome forgetting [6, 48, 60, 66].”
Representation Compensation Networks for Continual Semantic Segmentation (2022, background): “Some work propose to learn better representations by combining selfsupervised learning for feature extractor [9, 87] and solving class imbalance [39, 46, 54, 99, 102].”
DyTox: Transformers for Continual Learning with DYnamic TOken eXpansion (2021, methodology): “Performances are slightly lower than DyTox with SABs, however, they are still signiﬁcantly higher than previous state-of-the-art like WA [31], especially in “Last” accuracy.”
Online Continual Learning Via Candidates Voting (2021, methodology): “Recent studies [42, 45] also found that the bias of model weights towards new classes is one of the reasons for catastrophic forgetting.”
Class-Incremental Learning for Action Recognition in Videos (2021, neutral): “[46] rescale the weight vectors for the new classes by matching the average norm of the old weight vectors.”